EP3768150A1 - Procédé et système de détection d'un événement respiratoire d'un sujet et procédé de formation d'un modèle pour détecter un événement respiratoire - Google Patents
Procédé et système de détection d'un événement respiratoire d'un sujet et procédé de formation d'un modèle pour détecter un événement respiratoireInfo
- Publication number
- EP3768150A1 EP3768150A1 EP19709952.6A EP19709952A EP3768150A1 EP 3768150 A1 EP3768150 A1 EP 3768150A1 EP 19709952 A EP19709952 A EP 19709952A EP 3768150 A1 EP3768150 A1 EP 3768150A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- bio
- time
- impedance measurement
- sequences
- measurement signal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000000241 respiratory effect Effects 0.000 title claims abstract description 201
- 238000000034 method Methods 0.000 title claims abstract description 58
- 238000002847 impedance measurement Methods 0.000 claims abstract description 180
- 238000010801 machine learning Methods 0.000 claims abstract description 96
- 238000012549 training Methods 0.000 claims abstract description 82
- 230000001419 dependent effect Effects 0.000 claims abstract description 10
- 238000005259 measurement Methods 0.000 claims description 72
- 238000012545 processing Methods 0.000 claims description 45
- 238000007781 pre-processing Methods 0.000 claims description 20
- 206010021079 Hypopnoea Diseases 0.000 claims description 14
- 238000004891 communication Methods 0.000 claims description 14
- 208000003417 Central Sleep Apnea Diseases 0.000 claims description 8
- 238000012360 testing method Methods 0.000 claims description 8
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 208000001797 obstructive sleep apnea Diseases 0.000 claims description 7
- 230000015654 memory Effects 0.000 claims description 6
- 230000001360 synchronised effect Effects 0.000 claims description 4
- 238000001514 detection method Methods 0.000 description 28
- 208000008784 apnea Diseases 0.000 description 16
- 210000000038 chest Anatomy 0.000 description 12
- 201000002859 sleep apnea Diseases 0.000 description 12
- 230000029058 respiratory gaseous exchange Effects 0.000 description 10
- 230000002123 temporal effect Effects 0.000 description 9
- 238000012544 monitoring process Methods 0.000 description 8
- 230000000747 cardiac effect Effects 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000002347 injection Methods 0.000 description 5
- 239000007924 injection Substances 0.000 description 5
- 230000000875 corresponding effect Effects 0.000 description 4
- 238000013186 photoplethysmography Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 210000004072 lung Anatomy 0.000 description 3
- 230000000414 obstructive effect Effects 0.000 description 3
- 238000012805 post-processing Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000004622 sleep time Effects 0.000 description 3
- 230000001133 acceleration Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000036772 blood pressure Effects 0.000 description 2
- 230000036760 body temperature Effects 0.000 description 2
- 230000004069 differentiation Effects 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000002685 pulmonary effect Effects 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000036387 respiratory rate Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 206010019280 Heart failures Diseases 0.000 description 1
- 206010020772 Hypertension Diseases 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 239000000853 adhesive Substances 0.000 description 1
- 230000001070 adhesive effect Effects 0.000 description 1
- 230000002547 anomalous effect Effects 0.000 description 1
- 206010003119 arrhythmia Diseases 0.000 description 1
- 230000006793 arrhythmia Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 208000035475 disorder Diseases 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 229940079593 drug Drugs 0.000 description 1
- 238000002565 electrocardiography Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000000004 hemodynamic effect Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000010363 phase shift Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 208000019116 sleep disease Diseases 0.000 description 1
- 239000004753 textile Substances 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/48—Other medical applications
- A61B5/4806—Sleep evaluation
- A61B5/4818—Sleep apnoea
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/08—Detecting, measuring or recording devices for evaluating the respiratory organs
- A61B5/0809—Detecting, measuring or recording devices for evaluating the respiratory organs by impedance pneumography
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7271—Specific aspects of physiological measurement analysis
- A61B5/7275—Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/20—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for electronic clinical trials or questionnaires
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H40/00—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices
- G16H40/60—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices for the operation of medical equipment or devices
- G16H40/67—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices for the operation of medical equipment or devices for remote operation
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B2562/00—Details of sensors; Constructional details of sensor housings or probes; Accessories for sensors
- A61B2562/02—Details of sensors specially adapted for in-vivo measurements
- A61B2562/0219—Inertial sensors, e.g. accelerometers, gyroscopes, tilt switches
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/0205—Simultaneously evaluating both cardiovascular conditions and different types of body conditions, e.g. heart and respiratory condition
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/024—Detecting, measuring or recording pulse rate or heart rate
- A61B5/02416—Detecting, measuring or recording pulse rate or heart rate using photoplethysmograph signals, e.g. generated by infrared radiation
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
Definitions
- the present inventive concept relates to a method and a system for detecting a respiratory event of a subject.
- the present inventive concept also relates to a training method for forming a model for detecting a respiratory event.
- the present inventive concept relates to detection of respiratory events based on a bio-impedance measurement signal.
- Sleep apnea is one of the most common sleep-related disorders. Many cases of sleep apnea are undiagnosed as patients are rarely aware of their condition. The consequences of sleep apnea are, however, all too severe ranging from increased blood pressure, strokes and heart arrhythmia to critical heart failure.
- PSG polysomnography
- Bio-impedance signals are of increasing interest to use for monitoring of health of a subject.
- a bio-impedance signal may be modulated e.g. by breathing of a subject, and the bio-impedance signal may therefore be used for respiratory monitoring of the subject. This could be used for instance in sleep monitoring applications.
- Bio-impedance measurements may be performed with relatively simple equipment causing minimal or at least low inconvenience to the subject on which the bio-impedance measurements are performed. Therefore, use of bio-impedance measurements represent an interesting option in any type of monitoring, analysis or assessment of respiration, e.g. in screening of sleep apnea.
- WO 2004/1 12606 discloses an apparatus that includes means for acquisition of one or more bio-impedance signals, preferably across the thorax providing trans-thoracic measurements and across the neck region providing trans-cervical measurements; means for calculating components of the bio-impedance signals which can be ascribed to cardiac activity or to respiratory effort, means for obtaining measurements from the respiratory effort bio-impedance signal to produce diagnostic measures relating to obstructive and central apneas; and means for obtaining measurements from the cardiac bio-impedance signal which can provide a diagnostic measure of the impact of sleep disordered breathing on cardiac hemodynamic
- the apparatus makes use of two separate bio-impedance measurements, which should be detected across the thorax and across the neck region of a patient.
- bio impedance measurement signal is to be acquired.
- An objective of the present inventive concept is to improve respiratory monitoring using a bio-impedance signal.
- a particular objective of the present inventive concept is to provide a reliable detection of respiratory events.
- a method for detecting a respiratory event of a subject comprising: receiving a bio impedance measurement signal from the subject, said bio-impedance measurement signal being dependent on respiratory action of the subject; extracting at least one time-sequence of the bio-impedance measurement signal; and for each of the extracted time-sequences: comparing the bio impedance measurement signal with each of a plurality of machine learning models in an ensemble of machine learning models so as to form a set of predictions of occurrence of a respiratory event, wherein each prediction in the set is based on comparing the bio-impedance measurement signal with one machine learning model, wherein each model correlates features of time- sequences of a bio-impedance measurement signal with presence of a respiratory event in the time-sequences of the bio-impedance measurement signal and wherein each model is trained on a unique data set of training time-sequences; and deciding whether a respiratory event occurs in the extracted time-sequence
- the bio-impedance measurement signal need not be processed for extracting human engineered features, which would supposedly indicate respiratory events. Such human engineered features may potentially miss important markers of respiratory events.
- a bio-impedance measurement signal is compared to machine learning models for determining whether a respiratory event occurs.
- the machine learning model may in training of the model identify features in the bio-impedance measurement signal that are correlated to respiratory events, such that use of the machine learning models may allow making full use of information in the bio-impedance measurement signal for detecting respiratory events.
- the plurality of machine learning models allows determination of a set of predictions, such that the plurality of machine learning models contribute to detection of respiratory events and the detection of respiratory events may be based on an aggregated score from a plurality of models.
- the detection of respiratory events is less sensitive to a result based on a single machine learning model.
- the models may identify different correlations between features in the bio-impedance measurement signal and respiratory events. This further contributes to the detection of respiratory events being robust and that respiratory events may be reliably detected.
- the present inventive concept allows use of a bio-impedance measurement signal for robust detection of respiratory events.
- Bio-impedance measurements may be performed using a relatively simple equipment, which may cause no or insignificant inconvenience to the subject being tested. This implies that acquisition of signals for detecting respiratory events will not affect sleep of the subject such that high quality measurements may be performed while the subject is sleeping.
- the use of plurality of machine learning models enables robust detection of respiratory events even though a relatively simple bio-impedance measurement is made.
- the bio-impedance measurement signal being dependent on respiratory action of the subject may for instance be acquired to measure a bio-impedance on or across a thorax of the subject.
- a bio-impedance measurement is performed based on electrodes arranged on a thorax of a subject, chest expansion may cause a change in a current path between the electrodes, such that the bio-impedance is changed in relation to a respiratory effort.
- air has a different impedance than tissue. As an amount of air present in the lungs varies during a respiratory cycle, the bio-impedance is also changed in relation to respiratory airflow.
- a bio-impedance measurement signal which is dependent on respiratory action of the subject may be acquired in other manners, such as arranging electrodes to measure a bio impedance in any relation to lungs or airways of the subject, such as in relation to a neck portion of the subject.
- a respiratory event may be any event relating to a (temporally) anomalous respiration of the subject. This may include apnea and hypopnea, such as obstructive sleep apnea (OSA), central apnea (CA), obstructive hypopnea, and central hypopnea, and combinations thereof.
- OSA obstructive sleep apnea
- CA central apnea
- hypopnea central hypopnea
- the detection of a respiratory event may classify the type of respiratory event or may alternatively merely indicate that any respiratory event occurs, without classifying the respiratory event.
- the term“unique data set” as used herein should be construed as two models not being trained on identical data sets of training time-sequences.
- the two data sets used in training of the two models may share one or more training time-sequences, such that a particular training time-sequence may be present in both data sets.
- each data set will also include a training time-sequence which is not present in the other data set.
- more than two models may be trained on“unique data sets”, such that each model is trained on a data set which is not identical to any data set on which another model is trained.
- the bio-impedance measurement signal is compared with each of a plurality of machine learning models in an ensemble of machine learning models formed by neural networks with Long- Short Term Memory (LSTM) cells analyzing the unique data sets of training time-sequences.
- LSTM Long- Short Term Memory
- LSTM cells allow the machine learning models to model temporal aspects in the training time-sequences which enables the machine learning models to extract relevant temporal features and interactions in order to detect respiratory events.
- LSTM cells provide a powerful tool for extracting and using temporal information.
- the bio-impedance measurement signal is compared with each of a plurality of machine learning models in an ensemble of machine learning models formed by training each model on a unique data set that has been formed by balanced bootstrapping of acquired training time-sequences to combine positive time-sequences including a respiratory event, re-used between different data sets, with negative time- sequences not including a respiratory event, wherein at least a plurality of the negative time-sequences are dis-joint between different data sets.
- time-sequences corresponding to no respiratory event are far more common than time-sequences
- a straightforward option to balance a data set is to subsample negative training time-sequences leading to an equal amount of positive and negative training time-sequences. However, this also leads to a large amount of lost information.
- Another option is to oversample the positive training time-sequences, known as upsampling, leading to an equal amount of positive and negative training time-sequences.
- upsampling leading to an equal amount of positive and negative training time-sequences.
- the training time-sequences may be divided into unique data sets such that each data set is well-balanced between positive and negative training time-sequences.
- the positive time-sequences may be re-used between different data sets, such that each data set comprises several positive time-sequences, whereas the negative time-sequences may be arranged in separate data sets. In this way, each data set may have a good balance between positive and negative training time-sequences, while information is not discarded in creation of the models.
- the negative time-sequences i.e. all negative time-sequences, may be dis-joint between different data sets. This would imply that a particular negative time-sequence is only used once in the data sets. This may be a preferred manner of achieving the balanced bootstrapping.
- a plurality of negative time-sequences may be dis-joint between different data sets, while one or more negative time-sequences may be re used in two or more data sets.
- each of a plurality of machine learning models in an ensemble of machine learning models is formed by training each model on a unique data set defined by balanced bootstrapping, the machine learning models may be formed while taking into account all training data in a balanced manner. This facilitates that the machine learning models are well-trained and that the detection of respiratory events may be performed in a robust manner.
- the data set of training time-sequences have been acquired from the subject for whom respiratory events are detected.
- An individualized model may be advantageously used in long-term monitoring of a subject, e.g. for follow-up of a subject over long time periods.
- the machine learning models may be trained on general training time-sequences, which do not include or do not merely include training time-sequences of the subject for whom respiratory event detection is later to be performed.
- the machine learning models may be trained on large amounts of training time-sequences acquired from a large number of different patients. This may enable the machine learning models to be robust and enable generally detecting respiratory events for any subject.
- the machine learning model may be directly used in detection of respiratory events for a subject who has never been involved in training of the machine learning model.
- the machine learning model may be quickly used for detection of respiratory events of new patients, e.g. for diagnosing sleep apnea.
- the method further comprises receiving an additional measurement signal; extracting at least one time-sequence of the additional measurement signal, wherein the at least one time-sequence of the additional measurement signal corresponds to a simultaneous time period with the at least one time-sequence of the bio-impedance measurement signal; and for each of the extracted time-sequences: comparing the bio impedance measurement signal and the additional measurement signal with each of a plurality of machine learning models in an ensemble of machine learning models so as to form predictions being included in the set of predictions of occurrence of a respiratory event.
- the method may complement detection of respiratory events using bio-impedance measurement signals with at least one additional measurement.
- the additional measurement signal may be a measurement of a characteristic of the subject.
- the measurement signal may directly or indirectly relate to a vital sign and may provide additional information which may be used in detection of respiratory events.
- the additional measurement signal may be separately compared with a plurality of machine learning models in the ensemble of machine learning models, such that some machine learning models form predictions based on only the bio-impedance measurement signal and some machine learning models form predictions on only the additional measurement signal.
- machine learning models may form predictions based on both the bio-impedance measurement signal and the additional measurement signal.
- the use of the additional measurement signal may improve detection of respiratory events, since further information is acquired from the subject. Thus, using additional measurement signal(s), detection of respiratory events may be performed in an even more robust manner. However, it should be realized that an additional measurement signal is not necessary for enabling detection of respiratory events.
- the additional measurement signal may be any type of measurement which directly or indirectly provides a measure of e.g. a body temperature of the subject, a heart rate or pulse of the subject, a respiratory rate of the subject or a blood pressure of the subject.
- the additional measurement signal may for instance be a temperature signal, an electrocardiogram (ECG) signal, or a photoplethysmography (PPG) signal, whereby a direct measurement of a vital sign may be provided.
- ECG electrocardiogram
- PPG photoplethysmography
- the additional measurement signal may alternatively be an accelerometer signal, which may indicate movement of a body part (such as the chest) so as to indirectly provide information of a vital sign as the movement may represent cardiac or pulmonary activity.
- the method may comprise receiving a plurality (at least two) additional measurement signals, which may contribute to robust detection of respiratory events.
- the method further comprises pre processing the received bio-impedance measurement signal.
- the pre-processing of the received bio-impedance measurement signal may ensure that the bio-impedance measurement signal is clean and that the bio-impedance measurement signal represents a respiratory signal. Thus, pre-processing of the received bio-impedance measurement signal may provide an appropriate input for detection of respiratory events.
- the pre processing of the received bio-impedance measurement signal may be considered as transforming of the bio-impedance measurement signal into a signal representing respiration of the subject.
- the pre-processing of the received bio-impedance measurement signal may include subsampling of an acquired bio-impedance measurement signal.
- the pre-processing of the received bio-impedance measurement signal may include extracting a magnitude of a complex-valued signal.
- the pre-processing of the received bio-impedance measurement signal may include filtering of the signal, e.g. low-pass filtering of the signal, in order to extract respiratory information.
- the low-pass filter may be a
- the pre-processing of the received bio-impedance measurement signal may include removing motion artefacts from the signal.
- One or more of the above pre-processing operations may be any one or more of the above pre-processing operations.
- the pre-processing of the bio-impedance measurement signal may at least partially remove noise and may emphasize information in the bio-impedance measurement signal relating to respiration.
- the extracting of at least one time- sequence of the bio-impedance measurement signal comprises extracting a plurality of sequential time-sequences of the bio-impedance measurement signal.
- Detection of respiratory events may preferably be performed over a relatively long period of time. This may allow analysis of a condition of the subject, e.g. by detecting a frequency of occurrence of respiratory events.
- the received bio-impedance measurement signal may thus extend over a long time period and a plurality of sequential time-sequences may be extracted from the received bio-impedance measurement signal.
- Each time-sequence of the bio-impedance measurement signal may be separately analyzed in order to determine whether a respiratory event occurs within the time-sequence.
- the time-sequences may be partially overlapping. This may improve detection of respiratory events, since at least one time-sequence in a sequence of overlapping time-sequences may include a major portion of a duration of the respiratory event.
- the time-sequences are non-overlapping. This implies that a number of time-sequences may be small (compared to overlapping time-sequences) such that a process of detecting respiratory events may require smaller amount of time and computer resources.
- the method further comprises
- determining a measure of respiratory condition of the subject based on comparing a number of time-sequences, for which a decision that a
- the method may for instance determine a measure according to a standardized methodology, such that the measure of respiratory condition based on the method may be compared to measures obtained in other manners.
- the method may include determining an Apnea Hypopnea Index (AHI), which may be determined as a number of respiratory events divided by a total sleep time.
- AHI Apnea Hypopnea Index
- a relevant duration corresponding to sleep time of the subject may be determined.
- the received bio-impedance signal extends over a duration including a time period before the subject falls asleep or including time periods where the subject is awake, such time periods may be removed and not be part of the relevant duration period of the bio-impedance measurement signal.
- the comparing of the bio-impedance measurement signal with each of a plurality of machine learning models determines a plurality of sets of predictions for occurrences of different respiratory events, and wherein said deciding whether a respiratory event occurs is performed for each set of predictions, wherein the method further comprising outputting a prediction of occurrence of a respiratory event, wherein the outputting differentiates between respiratory events including central sleep apnea, obstructive sleep apnea and hypopnea.
- the differentiation between different types of respiratory events enables more detailed analysis of a condition of the subject, as types of respiratory events occurring may be taken into account in the analysis.
- Each model may be trained for enabling differentiation between different types of events.
- the training of the models for detecting different respiratory events may include balanced bootstrapping, by the models being trained on unique data sets which combine positive time-sequences for each type of respiratory event with negative time-sequences, wherein at least a plurality of the negative time-sequences are dis-joint between different data sets and positive time-sequences may be re-used between different data set.
- a method for forming a model for detecting a respiratory event of a subject comprising: receiving bio-impedance measurement signals from one or more subjects, said bio-impedance measurement signal being dependent on respiratory action of the subject; extracting a plurality of training time-sequences of the bio-impedance measurement signals; receiving information of occurrences of respiratory events, which information is synchronized with the plurality of training time-sequences so as to classify the training time-sequences as positive time-sequences representing a respiratory event or negative time- sequences not representing a respiratory event; determining a plurality of machine learning models to form an ensemble of machine learning models, wherein each model correlates features of time-sequences of a bio
- each model is trained on a unique data set of the plurality of training time- sequences.
- the training of the plurality of machine learning models to form an ensemble of machine learning models may allow respiratory events to be detected in a very robust manner based on the trained plurality of machine learning models.
- the models may identify different correlations between features in the bio-impedance measurement signal and respiratory events. This further contributes to the detection of respiratory events being robust and that respiratory events may be reliably detected.
- the information of occurrences of respiratory events may be received based on acquisition of another signal, from which respiratory events may be detected with large certainty.
- the other signal may be acquired in
- the received information of occurrences of respiratory events may be provided from a PSG system.
- the information of occurrences of respiratory events may be received from manual annotation, wherein time-sequence of bio-impedance measurement signals including a respiratory event may be tagged based on manual input.
- the determining of the plurality of machine learning models comprises using neural networks with Long-Short Term Memory (LSTM) cells for extracting a correlation of features of training time-sequences to presence of a respiratory event in the training time- sequences.
- LSTM Long-Short Term Memory
- LSTM cells allow the machine learning models to model temporal aspects in the training time-sequences which enables the machine learning models to extract relevant temporal features and interactions in order to detect respiratory events.
- LSTM cells provide a powerful tool for extracting and using temporal information.
- the method further comprises forming unique data sets to be used in training the plurality of machine learning models, wherein the forming of unique data sets comprises balanced bootstrapping of acquired training time-sequences to combine positive time- sequences including a respiratory event, re-used between different data sets, with negative time-sequences not including a respiratory event, wherein at least a plurality of the negative time-sequences are dis-joint between different data sets.
- the training time-sequences may be divided into unique data sets such that each data set is well-balanced between positive and negative training time-sequences.
- the positive time-sequences may be re-used between different data sets, such that each data set comprises several positive time-sequences, whereas the negative time-sequences may be arranged in separate data sets. In this way, each data set may have a good balance between positive and negative training time-sequences, while information is not discarded in creation of the models.
- each of a plurality of machine learning models in an ensemble of machine learning models is formed by training each model on a unique data set defined by balanced bootstrapping, the machine learning models may be formed while taking into account all training data in a balanced manner. This facilitates that the machine learning models are well- trained and that the detection of respiratory events may be performed in a robust manner.
- a system for detecting a respiratory event of a subject under test comprising: a bio impedance measurement sensor for acquiring a bio-impedance measurement signal from the subject, said bio-impedance measurement signal being dependent on respiratory action of the subject; and a processing unit, which is configured to: receive the bio-impedance measurement signal; extract at least one time-sequence of the bio-impedance measurement signal; and for each of the extracted time-sequences: compare the bio-impedance measurement signal with each of a plurality of machine learning models in an ensemble of machine learning models so as to form a set of predictions of occurrence of a respiratory event, wherein each prediction in the set is based on comparing the bio-impedance measurement signal with one machine learning model, wherein each model correlates features of time-sequences of a bio
- each model is trained on a unique data set of training time-sequences; and decide whether a respiratory event occurs in the extracted time-sequence based on the set of predictions.
- the system may thus comprise relatively simple equipment, including a bio-impedance measurement sensor, for acquiring of measurement signals from a subject in order to detect respiration events of the subject in a reliable and robust manner.
- the system may further comprise a processing unit for processing of the acquired bio-impedance measurement signal so that the system may decide whether respiratory event(s) occur in the acquired bio-impedance measurement signal.
- the bio-impedance measurement sensor may be configured to be arranged in relation to the subject for facilitating acquisition of the bio impedance measurement signal.
- the bio-impedance measurement sensor may be configured to be arranged in relation to the subject for facilitating acquisition of the bio impedance measurement signal.
- the measurement sensor may be arranged in a unit, which may be attachable to a body part of the subject for arranging the bio-impedance measurement sensor in contact with skin of the subject.
- the bio-impedance measurement sensor may be arranged in a patch for attachment to skin of the subject or in a belt for being strapped around a body part, such as the thorax, of the subject.
- the system further comprises a housing, wherein the bio-impedance measurement sensor and the processing unit are arranged in the housing.
- system may be embedded in a single physical unit, which provides a small form factor of the system. Also, all parts of the system in a single housing may be set up in relation to each other such that no set-up of the system may be necessary by a user before taking the system into use.
- the system further comprises a housing, wherein the bio-impedance measurement sensor is arranged in the housing, wherein the system further comprises a communication unit arranged in the housing, wherein the communication unit is configured to transmit the bio-impedance measurement signal to the processing unit.
- the system may thus comprise a plurality of physical units.
- the processing unit may be arranged in one of these units, which may be attached to or worn by the subject.
- the processing unit may be arranged in a central unit, which may be arranged to be mounted in close relation to a measurement setting, such as mounted to a bed frame, or which may simply be arranged to be positioned nearby, such as on a bed stand.
- the communication of the bio-impedance measurement signal to the processing unit may be performed through wired or wireless communication.
- the communication unit may be configured to communicate the bio-impedance measurement signal, e.g. via a telecommunication or computer network, to a processing unit, which may be arranged in a remote location.
- the processing unit may even be arranged“in the cloud”.
- the system may further comprise a storage unit for storing bio-impedance measurements.
- the system may be configured to acquire bio-impedance measurement signals from the subject and may store the bio-impedance measurements in the storage unit. Then, the bio-impedance measurements may be transferred at a later time to a processing unit for processing of the bio-impedance measurements.
- the storage unit may be arranged in a common housing with the bio impedance measurement sensor.
- the bio-impedance measurements may be stored in a housing in which the bio-impedance measurement sensor is arranged.
- the housing need not necessarily comprise any communication unit for communicating bio-impedance measurements, which may imply that the housing may be very compact and which may also imply that costs of a unit that is to be worn by the subject may be kept down.
- the bio-impedance measurements in the storage unit may be transferred to a processing unit, e.g. by physically connecting the storage unit to an external processing unit.
- bio-impedance measurements need not necessarily be performed in real time. Rather, the bio-impedance measurements may be stored and post-processing of bio impedance measurement signals acquried over a period of time may be performed.
- Fig. 1 is a schematic illustration of a system for respiratory monitoring according to an embodiment.
- Fig. 2 is a schematic flow chart of a training phase method for generating a plurality of machine learning models for detecting a respiratory event.
- Fig. 3 is a schematic flow chart of a deployment phase method for detecting respiratory events.
- Figs 4a-c is a schematic view of a balanced bootstrapping method for generating balanced data sets.
- Fig. 5 is a flow chart of a Long-Short Term Memory (LSTM) cell. Detailed description
- Fig. 1 illustrates a system 100 for respiratory monitoring of a subject.
- the system 100 may be configured to generate a current signal S1 that is to be applied to a subject and may comprise a bio-impedance measurement sensor 1 10 to acquire a bio-impedance measurement signal S2 providing information of the bio-impedance of the subject, which may be further processed for analyzing respiration of the subject.
- the bio-impedance measurement signal S2 may be provided to a processing unit 120 which may process the bio impedance measurement signal S2 to analyze the signal.
- the system 100 may comprise a current signal injection module 1 12.
- the current signal injection module 1 12 may be configured to generate and output the current signal S1 , which is to be applied to the subject.
- the current signal injection module 1 12 may comprise a current source for generating a current signal S1.
- the current signal injection module 1 12 may be configured to output an AC current signal.
- the system 100 may further comprise a bio-impedance measurement sensor 1 10.
- the bio-impedance measurement sensor 1 10 may be configured to receive voltage input signals representing a voltage generated by the current signal S1 applied to the subject.
- the bio-impedance measurement sensor 1 10 may be configured to extract a bio-impedance measurement signal S2 from the received voltage input signals.
- the bio-impedance measurement sensor 1 10 may comprise two or more electrodes 1 14, which may be arranged to be in contact with skin of the subject.
- the electrodes 1 14 may be connected to the current signal injection module 1 12 to receive the current signal S1 and provide the current signal through tissue of the subject.
- the electrodes 1 14 may also be connected to the bio-impedance measurement sensor 1 10 for providing voltage input signals that may be used for acquiring the bio-impedance measurement signal S2.
- the electrodes 1 14 may be arranged in a bipolar arrangement, wherein the same electrodes 1 14 are used for providing the current signal S1 to the subject and for acquiring the voltage input signals.
- the electrodes 114 may alternatively be arranged in a tetrapolar arrangement, wherein two electrodes are used for providing the current signal S1 to the subject and two other electrodes are used for acquiring the voltage input signals.
- More than two (or four) electrodes 114 may be provided, which may allow selection of which electrodes 1 14 to be used in a measurement, so that electrodes 114 providing highest quality bio-impedance measurement signal S2 may be selected.
- the selection of which electrodes 1 14 to be used may be performed in set-up of the system 100 or may be dynamically changed during signal acquisition e.g. when conditions for acquiring the bio-impedance signal change.
- the bio-impedance measurement sensor 1 10 with electrodes 1 14 may be configured to be attached on a thorax region of the subject.
- the bio impedance measurement sensor 1 10 may be arranged on a carrier 1 16 configured for being arranged on a thorax region of the subject, wherein the electrodes 114 may be mounted to be exposed on the carrier 1 16, such that the electrodes 1 14 may be arranged in contact with the skin of the subject.
- the carrier 1 16 may for instance comprise an adhesive patch, a textile/garment being worn by the subject, or a belt, which may be configured to be attached around the torso of the subject.
- the bio-impedance measurement sensor 1 10 may be configured for acquisition of a bio-impedance signal S2 which holds information of respiratory activity of the subject.
- the processing unit 120 may be configured to receive the bio impedance measurement signal S2 from the bio-impedance measurement sensor 1 10.
- the processing unit 120 may be configured to process the bio impedance measurement signal S2 for detecting occurrences of respiratory events.
- the processing unit 120 may be implemented in hardware, or as any combination of software and hardware.
- the processing unit 120 may, for instance, be implemented as software being executed on a general-purpose computer.
- the system 100 may thus comprise one or more physical processors, such as a central processing unit (CPU), which may execute the instructions of one or more computer programs in order to implement functionality of the processing unit 120.
- the system 120 may comprise a single processing unit, which may provide a plurality of functionalities e.g. as separate threads within the processing unit 120.
- the processing unit 120 may alternatively be implemented as firmware arranged e.g. in an embedded system, or as a specifically designed processing unit, such as an Application-Specific Integrated Circuit (ASIC) or a Field-Programmable Gate Array (FPGA).
- ASIC Application-Specific Integrated Circuit
- FPGA Field-Programmable Gate Array
- the system 100 may further comprise at least one additional measurement sensor 130.
- the additional measurement sensor 130 may be configured to measure e.g. a body temperature of the subject, a heart rate or pulse of the subject, a respiratory rate of the subject or a blood pressure of the subject.
- the additional measurement sensor 130 may for instance be a temperature sensor, which may be attached to skin of the subject, an electrocardiogram (ECG) sensor comprising electrodes for detecting an ECG, a photoplethysmograph (PPG) sensor comprising light source(s) and light sensor(s) for detecting a PPG, or an accelerometer, which may be configured to detect a movement of a body part (such as the chest), which may represent cardiac or pulmonary activity.
- ECG electrocardiogram
- PPG photoplethysmograph
- accelerometer which may be configured to detect a movement of a body part (such as the chest), which may represent cardiac or pulmonary activity.
- An additional measurement signal S3 may be acquired by the additional measurement sensor 130.
- the additional measurement signal S3 may be used by the processing unit 120, in combination with the bio impedance measurement signal S2 for detecting occurrence(s) of respiratory events.
- the system 100 may comprise one or more housings, in which the bio impedance measurement sensor 1 10, the processing unit 120 and the at least one additional measurement sensor 130 may be arranged.
- the housings may be connected by wires for allowing communication between the sensors 1 10, 130 and the processing unit 120.
- one or more of the sensors 1 10, 130 and the processing unit 120 may be set up for wireless communication.
- the system 100 may thus be delivered to be ready to use, e.g. in a single package with all parts of the system 100 already set up to communicate with each other.
- the processing unit 120 may be arranged in a housing on the carrier 1 16.
- the at least one additional measurement sensor 130 may also be arranged on the same carrier 1 16.
- the processing unit 120 may be arranged in a central housing, which may be separate from the carrier 1 16.
- the central housing may be configured to be mounted or arranged in a measurement setting, e.g. at a bed stand.
- the central housing may comprise a display, such that measurements and analysis results acquired by the system 100 may be displayed to a user.
- the system 100 may further comprise a communication unit 140 arranged in the carrier 1 16 or in a central housing.
- the communication unit 140 may be configured for communication via a telecommunication or computer network. This implies that the signals S2, S3 acquired by the system 100 may be communicated to a processing unit 120, which may be arranged in a remote location. Thus, the processing unit 120 may even be arranged“in the cloud”.
- the system 100 may further comprise a storage unit in which bio impedance measurements and additional measurements may be stored.
- the acquired bio-impedance measurements and additional measurements may thus be acquired for a period of time and the measurement results acquired during the period of time may be stored in the system 100, e.g. in a storage unit which may be arranged in the housing on the carrier 1 16.
- measurement results may be transferred from the storage unit to a
- the processing unit for post-processing of the measurement results any time after the measurement results have been acquired. This implies that the system 100 does not necessarily comprise a communication unit, since the
- measurement results may be transferred from the storage unit by physically connecting the storage unit to the processing unit or an intermediate unit, which may transfer the measurement results to the processing unit.
- the bio-impedance measurements and the additional measurements may also be separately acquired and even stored in separate storage units. The bio-impedance measurements and the additional measurements may then be synchronized in time during post-processing of the measurement results.
- the system 100 comprises relatively simple and inexpensive
- system 100 may be used for acquiring data over large time periods, such as acquiring data over multiple nights when analyzing sleep disorders, since the system 100 may be used in the home of the patient and a patient does not need to spend time at a laboratory, with limited availability.
- large amounts of data may be gathered for a subject, which may facilitate making better analysis of a subject’s condition.
- the processing unit 120 may be configured to extract at least one time- sequence of the bio-impedance measurement signal S2.
- a bio-impedance measurement signal S2 may be divided into a sequence of time-sequences, wherein sequential time-sequences may possibly be overlapping.
- the time- sequence may be analyzed in order to determine whether a respiratory event occurs within the time-sequence. By analyzing a plurality of time-sequences, it may then be possible to determine frequency of respiratory events and analyze a subject’s condition based on the frequency.
- a duration of a time-sequence may be selected to be within a range of e.g. 10-180 seconds. According to an embodiment, the duration may be selected to be within a range of 20-60 seconds. According to yet another embodiment, the duration may be selected to be 30 seconds.
- a duration of a time-sequence may be set sufficiently long to cover the duration of the respiratory event, but not too long so that a relative effect of the respiratory event may be hidden within the time-sequence.
- the processing unit 120 may further be configured to extract at least one time-sequence of the one or more additional measurement signals S3.
- An extracted time-sequence of the bio-impedance measurement signal S2 and/or an extracted time-sequence of the one or more additional measurement signals S3 may be compared to a plurality of machine learning models in an ensemble of machine learning models so as to form a set of predictions of occurrence of a respiratory event in the extracted time- sequence.
- the processing unit 120 may thus access an ensemble of machine learning models, which may e.g. be stored in a memory associated with the processing unit 120, so as to compare the extracted time-sequence with the stored plurality of machine learning models.
- the plurality of machine learning models may be configured to correlate features of time-sequences of a bio-impedance measurement signal S2 and/or features of time-sequences of one or more additional measurement signals S3 with presence of a respiratory event.
- Each model may be trained on a unique data set of training time-sequences, such that each model provides a separately trained manner of identifying respiratory events.
- the processing of the extracted time-sequence further comprises deciding whether a respiratory event occurs in the extracted time-sequence based on the set of predictions. Thanks to the use of a plurality of machine learning models, the detecting of respiratory events is robust, while using relatively simple measurements on the subject.
- a training phase method for generating a plurality of machine learning models will be described.
- the training phase method may comprise collecting 202 data from subjects.
- the collecting of data may comprise acquisition of a bio-impedance measurement signal S2 and, possibly, one or more additional measurement signals S3.
- a reference signal which may provide a reliable indication of respiratory events is acquired.
- PSG signals may be acquired, which may require complex and expensive hardware, but may provide a reliable detection of respiratory events.
- the reference signal may be analyzed for annotating 204 the reference signal to indicate respiratory events.
- the analysis of reference signal may be performed using a reliable technology for automated analysis of the reference signal.
- the analysis of reference signal may be performed by manual annotation, e.g. by manual annotation of PSG data using official guidelines, such as American Academy of Sleep Medicine (AASM)
- AASM American Academy of Sleep Medicine
- the annotation of reference signal may provide indication of time points when respiratory events occurs.
- Each respiratory event may further be classified based on a type of respiratory event, such as apnea and hypopnea, such as obstructive sleep apnea (OSA), central apnea (CA), obstructive hypopnea, and central hypopnea, and combinations thereof.
- the acquired signals i.e. the bio-impedance measurement signal S2, the one or more additional measurement signals S3, and the reference signal, may be time synchronized 206 to ensure that the annotation of the reference signal may be correctly applied to the information in the bio-impedance measurement signal S2 and the one or more additional measurement signals S3.
- the time synchronization may be performed before or after the annotation of the reference signal.
- One or more of the acquired signals i.e. the bio-impedance
- the measurement signal S2 the one or more additional measurement signals S3, and the reference signal, may further be preprocessed 208.
- preprocessing of signals may be used for noise removal and extracting relevant respiratory information in a signal which may include information associated with other sources, such as cardiac activity.
- the bio-impedance measurement signal S2 and the one or more additional measurement signals S3 may then be partitioned 210 in a sequence of time-sequences (or epochs).
- the epochs may be overlapping.
- Each epoch may be labeled based on the annotation of the reference signal, such that, for each epoch there is an indication whether a respiratory event occurs, e.g. a“0” indicating that no respiratory event is occurring and a “1” indicating that a respiratory event is occurring.
- the labelling may be performed generally for any respiratory event occurring or may be performed for each type of respiratory event, such that a plurality of labels (0/1 ) may be set for a plurality of types of respiratory events.
- the epochs are to be used as training time-sequences for training machine learning models.
- a collected data set will comprise a much larger number of negative epochs (no respiratory event occurring) than positive epochs (respiratory event
- the epochs are arranged 212 in a plurality of training data sets to handle the imbalance through balanced bootstrapping.
- the positive epochs are re-used between different data sets, whereas at least a plurality of negative epochs are dis-joint between different data sets. In this manner, all epochs may be used, while, for each data set, there is a good balance between positive and negative epochs.
- the plurality of training data sets is then used to calculate 214 a plurality of machine learning models.
- Each model may be trained using one of the data sets.
- LSTM Long-Short Term Memory
- LSTM Long-Short Term Memory
- GRU Gated Recurrent Unit
- RNN Recurrent Neural Network
- Each model may be trained to generate a prediction of occurrence of a respiratory event.
- the plurality of machine learning models may further be used for averaging predictions of individual models into an aggregate model that may output a single score (e.g. a mean of probabilities).
- the training phase method may include specifying 216 a decision threshold based on a set of predictions for enabling a decision whether a respiratory event occurs based on whether an aggregate score of the machine learning models is above or below the decision threshold.
- Fig. 3 a deployment phase method for using the plurality of machine learning models in detecting of respiratory events will be described.
- the detection of respiratory events is made based on only the bio-impedance measurement signal S2, possibly in combination with the one or more additional measurement signals S3.
- the reference signal is no longer needed, as the system 100 has now been trained for detecting respiratory events based on the bio-impedance measurement signal S2, possibly in combination with the one or more additional measurement signals S3.
- the method comprises collecting 302 data from a subject under test by acquisition of the bio-impedance measurement signal S2, possibly in combination with the one or more additional measurement signals S3.
- the method may further comprise preprocessing 304 the bio impedance measurement signal S2 and the one or more additional measurement signals S3. Similar to the training phase method, the
- preprocessing of signals may be used for noise removal and extracting relevant respiratory information in a signal.
- the bio-impedance measurement signal S2 and the one or more additional measurement signals S3 may then be partitioned 306 in a sequence of time-sequences (or epochs).
- the epochs may be overlapping.
- Each epoch may then be analyzed in order to determine whether a respiratory event is occurring within the epoch.
- the analysis may comprise evaluating 308 the epoch using an aggregated machine learning model, as generated according to the above description.
- the epoch may be compared to each model in a plurality of machine learning models for determining a set of predictions of occurrence of a respiratory event within the epoch.
- the set of predictions may be aggregated into a single, combined score for a probability of a respiratory event occurring.
- a decision may be taken 310 whether a respiratory event occurs within the epoch.
- the decision may be based on determining whether the aggregated score obtained by the machine learning model exceeds a set threshold or not.
- a system 100 which is trained is configured to record three types of signals: bio-impedance (BioZ), ElectroCardioGraphy (ECG) and acceleration.
- Bio-impedance BioZ
- ECG ElectroCardioGraphy
- the bio-impedance is measured at 1024Hz
- the ECG is measured at 512Hz
- the acceleration consisting of 3 separate components, is measured at 32Hz.
- these signals are combined with the signals from the PSG device recorded at 200Hz.
- a significant frequency shift is visible after a while as sampling frequencies of the signals are not exact.
- the system 100 and the PSG device are attached to the patient, they may not be activated at the same time. Hence, accurate time- alignment of both devices may be needed.
- the ECG signal is used as it is available in both devices.
- the time alignment may be based on other information.
- these machine learning models are used to predict the occurrence of sleep apnea in data of new, unseen subjects, wherein the data is recorded using only the system 100.
- the preprocessing of bio-impedance measurement signals S2 may convert raw bio-impedance input signals into epochs comprising relevant information. A complex-valued bio-impedance measurement signal may thus be transformed into a respiratory signal.
- the preprocessing may include that the raw bio-impedance
- measurement signal is subsampled to a sampling frequency of 5Hz after which a magnitude of the complex-valued bio-impedance signal is extracted.
- features may be extracted in another manner from the bio impedance signal. For instance, just the real part of the complex-valued bio impedance signal may be extracted.
- the bio-impedance measurement signal in this case the bio-impedance magnitude (or real value of the bio- impedacne), may in preprocessing of the signal be passed through a filter, such as a low-pass Butterworth filter.
- the Butterworth filter may ensure that information in the bio-impedance measurement signal S2 relating to respiration of the subject is extracted in a robust manner from other information in the bio-impedance measurement signal S2, such as cardiac activity information.
- the Butterworth filter may be a fourth order Butterworth filter with cutoff frequency of 0.7Hz, but it should be realized that other settings may be used.
- the preprocessing may further include removing of motion artefacts from the filtered signal by applying a moving average filter to the filtered signal and subtracting the signal obtained from applying the moving average filter from the original filtered signal.
- the moving average filter may for instance have a width of 4 seconds.
- a full-length respiratory signal may be converted into 30 second epochs with a stride of 1 second in between each other, i.e. there is a major overlap between two sequential epochs (only 1 s of the signal is differing).
- each epoch may be standardized separately in a range of [-1 ; 1 ]. This may improve modelling capabilities of the machine learning models.
- simultaneous measurements are gathered from the system 100 as well as from a PSG device.
- the data from the PSG device may be used to create an apnea-hypopnea annotation by a trained nurse.
- an annotation is added based on the PSG annotations of the nurse. If at the end of the epoch the nurse annotated an apnea or hypopnea event, the epoch is labeled an apnea epoch.
- different respiratory events may be distinguished. However, in other embodiments, no distinction is made in OSA, CA or hypopnea.
- the epochs generated by the preprocessing normally comprise of a large majority (95%) of apnea-negative epochs and a small minority (5%) of apnea-positive epochs leading to an imbalanced data set.
- apnea-negative epochs normally comprise of a large majority (95%) of apnea-negative epochs and a small minority (5%) of apnea-positive epochs leading to an imbalanced data set.
- data balancing is advantageous in allowing a data set to be properly modelled.
- a balanced bootstrapping method is used. As illustrated in Figs 4a-c, with the balanced bootstrapping method, the large imbalanced data set 400, illustrated in Fig. 4a, is split up into several smaller balanced data sets 404a-g.
- the majority class (indicated by white areas in Figs 4a-c) in the unbalanced data set 400 is split into subsets 402a-g with size equal to the minority class (indicated by hatched areas in Figs 4a-c), as illustrated in Fig. 4b.
- the epochs of the minority class are appended to the different sets of the majority class leading to multiple balanced data sets 404a-g, as illustrated in Fig. 4c.
- Each data set may have all epochs from the minority class and a disjoint set of epochs from the majority class.
- Each of these data sets 404a-g can now be used to build a separate model.
- epochs from the minority class are used in each of the balanced data sets.
- An apnea-positive epoch may be re-used in several data sets 404a-g, but not necessarily re-used in every data set 404a-g.
- not necessarily all majority epochs are used in a single data set. Rather, one or more majority epochs may be re-used in the data sets 404a-g, such that the sets of epochs from the majority class are not completely dis-joint.
- Each of the data sets 404a-g generated in the balanced bootstrapping procedure is modelled individually.
- the data being modelled comprises a lot of temporal information
- neural networks with LSTM cells may be used as these are powerful tools for extracting and using temporal information.
- x represents the input sequence
- h represents the hidden sequence
- t represents a time point in the sequences
- c represents the cell vector and /
- f and o represent the input gate, forget gate and the output gate, respectively
- the s operator is the logistic sigmoid function
- the b terms are the bias terms
- the matrices W are weight matrices between two entities.
- the LSTM cells may automatically extract relevant feature data through time.
- the network architecture used comprises 50 base LSTM cells followed by a dropout layer with dropout probability 0.2. Adding a dropout layer allows for a more powerful generalization of the model to unseen data. After the dropout layer, 10 regular dense nodes are added followed by another dropout layer with dropout probability 0.2. Finally, a single dense neural network node is appended with sigmoid activation function for classification of epochs.
- the network may be trained with minibatches of 32 epochs consisting of 16 positive and 16 negative epochs using the binary crossentropy loss function which is optimized using the adadelta optimiser. Gradient clipping is applied to a value of 100.
- the balanced bootstrapping procedure has created several distinct data sets for models to be trained on.
- each of these models is trained individually. To test a new epoch, it may be tested with any of these models. However, due to random differences in the models and differences in the training data sets, these models will have slightly different predictions. As these predictions represent probabilities, we can take a mean of these predictions to get a mean probability of the output. The aggregated output from all models provides a more robust estimate of the true probability of apnea for that epoch.
- This aggregated estimate can be further fine-tuned to specific needs of an application for detecting respiratory events.
- an epoch may be determined as comprising an apnea event by determining if the probability of apnea is larger than 0.5.
- this decision threshold may be tuned by varying the decision threshold in the range [0; 1 ]. Based on such tuning, resulting sensitivity and specificity scores may be computed. A most beneficial decision threshold (largest sensitivity and specificity) may then be chosen as the final decision threshold value t.
- the Apnea Hypopnea Index is computed as the number of respiratory events divided by the total sleep time following the official AASM guidelines. As each epoch ends 1 second after the other, the method may provide a prediction of the probability of sleep apnea for each second of the signal. If the probability is larger than t, a particular second in the signal is flagged as a respiratory event. These generated annotations are then used to compute the AHI. Using this AHI score, patients may be classified with normal breathing, mild sleep apnea, moderate sleep apnea or severe sleep apnea.
Landscapes
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- Pathology (AREA)
- Artificial Intelligence (AREA)
- Animal Behavior & Ethology (AREA)
- Veterinary Medicine (AREA)
- Heart & Thoracic Surgery (AREA)
- Molecular Biology (AREA)
- Surgery (AREA)
- Biophysics (AREA)
- Physiology (AREA)
- Primary Health Care (AREA)
- Epidemiology (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Physics (AREA)
- Psychiatry (AREA)
- Signal Processing (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Fuzzy Systems (AREA)
- Pulmonology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Business, Economics & Management (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP18162591 | 2018-03-19 | ||
PCT/EP2019/056222 WO2019179836A1 (fr) | 2018-03-19 | 2019-03-13 | Procédé et système de détection d'un événement respiratoire d'un sujet et procédé de formation d'un modèle pour détecter un événement respiratoire |
Publications (1)
Publication Number | Publication Date |
---|---|
EP3768150A1 true EP3768150A1 (fr) | 2021-01-27 |
Family
ID=61800296
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP19709952.6A Pending EP3768150A1 (fr) | 2018-03-19 | 2019-03-13 | Procédé et système de détection d'un événement respiratoire d'un sujet et procédé de formation d'un modèle pour détecter un événement respiratoire |
Country Status (5)
Country | Link |
---|---|
US (1) | US11903724B2 (fr) |
EP (1) | EP3768150A1 (fr) |
JP (1) | JP2021518246A (fr) |
CA (1) | CA3094421A1 (fr) |
WO (1) | WO2019179836A1 (fr) |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11139048B2 (en) | 2017-07-18 | 2021-10-05 | Analytics For Life Inc. | Discovering novel features to use in machine learning techniques, such as machine learning techniques for diagnosing medical conditions |
US11062792B2 (en) | 2017-07-18 | 2021-07-13 | Analytics For Life Inc. | Discovering genomes to use in machine learning techniques |
CN111248916B (zh) * | 2020-01-03 | 2023-08-11 | 博博科技(苏州)有限公司 | Bcg信号的ahi指数检测方法 |
CA3089395A1 (fr) * | 2020-06-18 | 2021-12-18 | University Of Manitoba | Systemes et methodes pour verifier l`apnee du sommeil obstructive pendant l`etat eveille au moyen de renseignements anthropometriques et des bruits de respiration tracheaux |
US20220054039A1 (en) * | 2020-08-20 | 2022-02-24 | Samsung Electronics Co., Ltd. | Breathing measurement and management using an electronic device |
RU2766046C1 (ru) * | 2020-12-22 | 2022-02-07 | Федеральное государственное бюджетное образовательное учреждение высшего образования "Рязанский государственный радиотехнический университет имени В.Ф. Уткина" | Способ дистанционной регистрации процесса дыхания пациента и устройство для его осуществления |
CN113273992B (zh) * | 2021-05-11 | 2023-04-18 | 清华大学深圳国际研究生院 | 信号处理方法及装置 |
CN115804585B (zh) * | 2023-02-08 | 2023-06-13 | 浙江大学 | 一种基于机械通气波形检测气道高阻力的方法及系统 |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH10295695A (ja) * | 1997-04-30 | 1998-11-10 | Mitsubishi Chem Corp | 無呼吸検出器 |
US6641542B2 (en) * | 2001-04-30 | 2003-11-04 | Medtronic, Inc. | Method and apparatus to detect and treat sleep respiratory events |
JP3947379B2 (ja) * | 2001-10-12 | 2007-07-18 | 積水化学工業株式会社 | 電気特性測定装置 |
IES20030467A2 (en) | 2003-06-24 | 2005-02-09 | Univ Dublin | Methods for detecting sleep apnea using bioimpedance measurements |
US7276031B2 (en) | 2004-05-12 | 2007-10-02 | New York University | System and method for classifying patient's breathing using artificial neural network |
ATE476716T1 (de) * | 2004-11-19 | 2010-08-15 | Koninkl Philips Electronics Nv | Stratifikationsverfahren zur überwindung ungleichmässiger fallanzahl bei rechnergestützter lungenknoten-falsch-positiv-reduktion |
CN101815465B (zh) * | 2007-10-02 | 2014-04-23 | 康迪医疗革新有限公司 | 从心电图得出的呼吸暂停/呼吸减弱指数 |
US20130172759A1 (en) * | 2011-08-08 | 2013-07-04 | Richard J. Melker | Systems And Methods For Using Photoplethysmography In The Administration Of Narcotic Reversal Agents |
JP5694139B2 (ja) * | 2011-12-28 | 2015-04-01 | 日本光電工業株式会社 | 睡眠中における無呼吸低呼吸状態の検出装置 |
CA2877430C (fr) * | 2012-06-21 | 2021-07-06 | Philip Morris Products S.A. | Systemes et procedes pour generer des signatures de biomarqueurs au moyen d'ensembles doubles integres et de techniques d'annelage simulees |
WO2015059700A1 (fr) | 2013-10-24 | 2015-04-30 | Breathevision Ltd. | Système de surveillance de mouvement |
US10028665B2 (en) * | 2014-09-12 | 2018-07-24 | Panasonic Intellectual Property Management Co., Ltd. | Biosignal determining device and biosignal determining method |
WO2016093927A2 (fr) * | 2014-12-08 | 2016-06-16 | University Of Washington | Systèmes et procédés d'identification du mouvement d'un sujet |
JP2017045331A (ja) * | 2015-08-27 | 2017-03-02 | キヤノン株式会社 | 画像処理方法、画像処理装置、及びプログラム |
-
2019
- 2019-03-13 EP EP19709952.6A patent/EP3768150A1/fr active Pending
- 2019-03-13 CA CA3094421A patent/CA3094421A1/fr active Pending
- 2019-03-13 JP JP2021500335A patent/JP2021518246A/ja active Pending
- 2019-03-13 WO PCT/EP2019/056222 patent/WO2019179836A1/fr unknown
- 2019-03-13 US US16/981,931 patent/US11903724B2/en active Active
Also Published As
Publication number | Publication date |
---|---|
JP2021518246A (ja) | 2021-08-02 |
US20210030353A1 (en) | 2021-02-04 |
WO2019179836A1 (fr) | 2019-09-26 |
CA3094421A1 (fr) | 2019-09-26 |
US11903724B2 (en) | 2024-02-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11903724B2 (en) | Method and a system for detecting a respiratory event of a subject and a method for forming a model for detecting a respiratory event | |
US20240221948A1 (en) | Systems and methods of analyte measurement analysis | |
Pathinarupothi et al. | Instantaneous heart rate as a robust feature for sleep apnea severity detection using deep learning | |
CN101815465B (zh) | 从心电图得出的呼吸暂停/呼吸减弱指数 | |
WO2019161609A1 (fr) | Procédé d'analyse de données de surveillance multi-paramètres et moniteur multi-paramètres | |
Shin et al. | Automatic ballistocardiogram (BCG) beat detection using a template matching approach | |
EP3672474A1 (fr) | Procédé de détection d'anomalies dans des signaux ecg | |
US11783924B2 (en) | ECG information processing method and ECG workstation | |
US20210007618A1 (en) | Electrocardiogram information processing method and electrocardiogram workstation system | |
US8992435B2 (en) | System and method for classifying a heart sound | |
Klug et al. | The BeMoBIL Pipeline for automated analyses of multimodal mobile brain and body imaging data | |
CN108186000A (zh) | 基于心冲击信号与光电信号的实时血压监测系统及方法 | |
Chowdhury et al. | Machine learning in wearable biomedical systems | |
CN116194044A (zh) | 用于分析心律的方法和系统 | |
CN110022762A (zh) | 监测用户心脏活动的方法和系统 | |
Bourke et al. | Validation of the activPAL3 in free-living and laboratory scenarios for the measurement of physical activity, stepping, and transitions in older adults | |
Alivar et al. | Motion detection in bed-based ballistocardiogram to quantify sleep quality | |
Rossi et al. | Sleep-see-through: Explainable deep learning for sleep event detection and quantification from wearable somnography | |
Thommandram et al. | A rule-based temporal analysis method for online health analytics and its application for real-time detection of neonatal spells | |
CN104981207B (zh) | 用于心脏获取的同步生理学测量 | |
US20220330872A1 (en) | Bio-signal measuring apparatus for detecting abnormal signal section in electrocardiogram data by using heart sound data related to electrocardiogram data, and bio-signal measuring method | |
Heise et al. | Unobtrusively detecting apnea and hypopnea events via a hydraulic bed sensor | |
JP2020048622A (ja) | 生体状態推定装置 | |
CN116211251A (zh) | 基于消费级设备心率数据检测睡眠呼吸暂停的系统及设备 | |
KR102645586B1 (ko) | 생체신호를 이용한 수면 호흡 상태 분류 장치 및 방법 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
17P | Request for examination filed |
Effective date: 20201009 |
|
AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
AX | Request for extension of the european patent |
Extension state: BA ME |
|
DAV | Request for validation of the european patent (deleted) | ||
DAX | Request for extension of the european patent (deleted) | ||
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
17Q | First examination report despatched |
Effective date: 20240426 |